Traffic Light Detection for Autonomous Vehicles in Diverse Environments Through Transfer Learning

Traffic light detection for autonomous vehicles (AVs) is challenging in varying outdoor lighting conditions. While several studies show promise, handling variable lighting remains a research focus. This paper proposes a traffic light detection system using image processing and machine learning. The system trained multiple models separately with day and night data so that the proposed system may select the appropriate model using brightness histograms. We employed two versions of the YOLO v8 model, fine-tuned for day and night detection. Our approach was tested on a custom dataset of 470 images from Dhaka, Bangladesh and the LISA dataset. The system achieved perfect accuracy on the custom dataset and 98.99% on LISA for day or night classification. Object detection scores for mAP50 are 97.6% (day) and 83.1% (night) on the custom dataset and 92.6% (day) and 96.1% (night) on LISA. These results confirm the effectiveness of the proposed method in variable lighting conditions.